Re: Fix performance of generic atomics

On 2017-09-06 07:23, Tom Lane wrote:
> Jeff Janes <jeff.janes@gmail.com> writes:
>> What scale factor and client count? How many cores per socket?  It 
>> looks
>> like Sokolov was just starting to see gains at 200 clients on 72 
>> cores,
>> using -N transaction.
> This means that Sokolov's proposed changes in atomics/generic.h
> ought to be uninteresting for performance on any platform we care about 
> --- at
> least for pg_atomic_fetch_add_u32().

May I cite you this way:
"Tom Lane says PostgreSQL core team doesn't care for 4 socket 72 core
Intel Xeon servers"
???

To be honestly, I didn't measure exact impact of changing 
pg_atomic_fetch_add_u32.
I found issue with pg_atomic_fetch_or_u32, cause it is highly contended
both in LWLock, and in buffer page state management. I found changing
loop for generic pg_atomic_fetch_or_u32 already gives improvement.
And I changed loop for other generic atomic functions to make
all functions uniform.

It is bad style guide not to changing worse (but correct) code into
better (and also correct) just because you can't measure difference
(in my opinion. Perhaps i am mistaken).

(and yes: gcc intrinsic gives more improvement).

--
With regards,
Sokolov Yura aka funny_falcon
Postgres Professional: https://postgrespro.ru
The Russian Postgres Company